Master of Philosophy
School of Electrical, Computer and Telecommunications Engineering
Stephens-Fripp, Benjamin, Combining Local and Global Features in Automatic Affect Recognition from Body Posture and Gait, Master of Philosophy thesis, School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, 2016. https://ro.uow.edu.au/theses/4923
There has been a growing interest in machine-based recognition of emotions from body gait and posture, and its combination with other modalities. Applications such as human computer interaction, social robotics, and security have been the driving force behind such trend. The majority of the previous work in automatic affect perception deploys only either local features or global features. Whilst a combination of both types of features are deployed in applications such as object recognition and facial recognition, the literature does not reveal any study in affect recognition from body language using combined global and local features. In this thesis, such gap is addressed by examining how deploying a combination of local and global features can improve the recognition rate in automatic classification of emotions using gait and posture.
The motion data used in the study comprising kinematic parameters associated with the gait and posture of a number of actors expressing a set of emotions, were recorded electronically using an inertia motion capture system. A combination of local and global features proposed by Kapur et al. and Zacharatos et al., respectively, were used in the classification process using WEKA classification system. Additional global features of shape flow and shaping, horizontal and vertical symmetry were added to the combination feature set to increase the performance of the classifier.
The results obtained in the analysis demonstrate that deploying a combination of local and global features leads to a more robust and reliable method for automatic affect recognition from body language as it improves accuracy across a range of classifiers. This research also demonstrates that the inclusion of the additional features, which represent additional Laban Movement Analysis components, increases the maximum classification accuracy from 88.5% to 92.3%.
Achieving better automatic affect recognition rates can lead to increased application of the approach, improved usefulness and reliability of such systems.